Data centers: the US AI bet equals 3.6% of GDP spread over several years

The scale of the buildout goes beyond anything the American economy has seen. Economist Stijn Van Nieuwerburgh, in an analysis reported by heise online, calculates that AI infrastructure investment in the United States works out to 3.6% of the country's gross domestic product once spread over several years. That is almost double the previous record, set during the railway boom, when 2.24% of GDP at the time was committed.

A comparison with past major cycles

To measure the anomaly, the economist places the phenomenon in the history of large networks. The canal, electrification, highway construction and fiber optic booms, the last of them during the internet bubble, mobilized between 0.5% and 1.13% of GDP. The current buildout sits at a completely different level.

The sensitive point is not the amount but how it is paid for. Until 2024, the large groups concerned, Oracle, Amazon, Alphabet, Microsoft and Meta, took in at least twice what they spent. In 2026, that ratio should flip. Investments should clearly exceed revenues that are nonetheless rising sharply every year. With insufficient cash on hand, a growing share of the financing comes from outside, which increases the risk.

Financial models that keep getting more opaque

The economist draws a parallel with earlier episodes. A general-purpose technology drives complementary investment in a vast physical network. Returns depend on uncertain future demand, and the financing needs are considerable. One notable difference exists, though: the rapid obsolescence of the hardware bought in bulk. Infrastructure financing stretches over long periods, while graphics cards and other components potentially lose their value far faster. That creates a significant imbalance.

The analysis even evokes the subprime crisis that led to the 2008 financial crisis, while specifying that it does not suggest imminent financial difficulties. Meanwhile, commitments are multiplying. Anthropic announced it would spend 11.6 billion dollars over seven years on Akamai's cloud infrastructure, more than six times the amount of a 1.8 billion deal reported in May. The commitment is not unconditional, however. It depends on meeting certain delivery and availability requirements, and the two companies can terminate it under certain conditions.

Local constraints are mounting

At the local level, resistance is organizing. In Virginia, the Loudoun County board, in the historic heart of data centers, voted for a pause on new siting applications. The debate covers electricity, water and acceptability for nearby residents, all subjects moving up the political agenda.

A general-purpose technology drives investments in a physical network whose returns depend on uncertain future demand.